The battle for computing power infrastructure has only just kicked off in the liquid cooling track.

Orders are booked through 2027, revenue growth exceeds 400%, and production capacity cannot keep up with delivery... This is not the story of a large Internet company, but the 2026 status quo of the liquid cooling industry.

Driven continuously by the demand for high-density AI computing power, liquid cooling is undergoing a status leap: what used to be an optional supporting facility in data center construction is now evolving into a bottom-tier infrastructure that must be considered synchronously during the construction phase of high-density intelligent computing clusters.

The equipment end of the industrial chain has taken the lead in a round of concentrated explosion: leading manufacturers see surging orders, enterprises are expanding production on a large scale, and new cross-border players are pouring in one after another.

This "cool business" ignited by the boom in computing power construction is becoming hotter than ever.

From a supporting role to rigid demand, the computing power wave boosts the "cool business"

Many people's impression of data center temperature control still stays in the traditional air cooling era.

In the past IDC construction logic, heat dissipation was only a supporting link: servers and cabinets were the core, and air conditioners were supporting equipment purchased on demand in the later stage, with a high degree of standardization, and would not become a decisive constraint on the early architecture design of data centers.

AI computing power clusters have rewritten this rule: the power consumption of a single GPU computing chip continues to rise, and the power of high-density cabinets has broken through 30kW, 50kW or even higher. In such ultra-high-density scenarios, the heat dissipation solution of traditional air cooling has reached the ceiling of engineering implementation.

Not only is it difficult to meet the PUE indicator, but continuous high temperature will also affect the operation stability of chips, and even shorten the service life of hardware. When the computing power density rises to a certain threshold, heat dissipation is no longer an optional supporting facility, but a bottom-tier infrastructure that needs to be planned synchronously in the early construction stage of high-density intelligent computing centers.

Essentially, this status leap of liquid cooling is a structural demand driven by engineering constraints brought by high-density computing power hardware.

The market demand presents a clear stratification, and the differences in the demands of different types of customers are forming the basic market foundation of the liquid cooling market.

For example, leading cloud vendors and large intelligent computing cluster operators give top priority to long-term stable operation capability, are willing to invest a long cycle to verify the reliability of solutions, and cost usually ranks second; when commercial IDC service providers build new data centers, they will take into account both implementation cost and transformation compatibility, and give priority to solutions with higher maturity; for new computing power projects in various regions, it is necessary to quickly complete the implementation of computing power indicators, and they are more sensitive to the delivery cycle and one-time input cost.

The interweaving of multiple demands has jointly pushed up the heat of industry orders, and also made the market demand structure more complex.

What follows is the explosion of industry revenue and order data. On September 28, 2026, Shanghai Securities News disclosed that the liquid cooling orders for AI servers continue to surge, and many companies have begun to expand production. Many industrial chain manufacturers also recently stated that the scheduling of AI server liquid cooling orders has been extended to 2027, the revenue growth rate of some enterprises exceeds 400%, and the gap between production capacity and delivery continues to widen.

However, industry insiders remind that the order caliber disclosed externally needs to be carefully identified. Many orders in the liquid cooling industry are framework agreements or letters of intent, which only stipulate the cooperation intention and estimated procurement scale, and usually come with preconditions such as prototype testing and on-site pilot verification, and cannot be directly equated with recognizable revenue.

Letters of intent usually go through a long period of reliability testing, and then to the project implementation and construction, before they can be converted into actual delivery orders. However, in market promotion, different types of orders are often mixed together, amplifying the expectation of industry prosperity.

This also means that the increment of the liquid cooling track has real underlying support, but the order scale seen in the market itself contains a certain "expected moisture". The demand growth brought by computing power is not a false boom, but if you simply equate paper orders directly with future revenue, it is easy to misjudge the real situation of the industry.

Liquid cooling has transformed from a supporting role of data centers to a part of computing power infrastructure, and the underlying logic has changed, but the demand explosion is only one side of the story. A large amount of capital and new players are pouring into the track, and differences in technical routes and engineering implementation problems are gradually emerging.

No standard answer for routes, the battlefield differentiation between cold plate and immersion

As the industry heats up, discussions on the two technical routes of cold plate liquid cooling and immersion liquid cooling continue to ferment.

Public opinion fields often fall into the binary debate of "who will eventually replace who", but it is not difficult to find by going deep into the front line of the industry that leading computing power customers will not put all their bets on a single solution, and most of them are testing multiple technical paths in parallel.

The core of this route dispute is not to judge the quality of the technology, but to make trade-offs on cost, operation and maintenance, and reliability in different computing power scenarios. The two routes will most likely coexist for a long time, and it is difficult to form a winner-takes-all pattern.

Cold plate liquid cooling is the solution with the largest number of implemented projects at present. It installs a cold plate inside the server, relies on cooling liquid to take away the heat of the chip, and the external supporting pipeline completes heat exchange. Its core advantage is that it has limited changes to the existing server architecture, has better compatibility for the transformation of existing data centers, and has accumulated a lot of engineering implementation experience.

For IDC service providers and most new intelligent computing centers, the cold plate solution is a risk-controllable option. But its shortcomings are also obvious: there are a large number of pipeline joints, and the leakage risk points are scattered; heat exchange has an engineering boundary, and in the scenario of ultra-high-density cabinets, the heat dissipation capacity will gradually face pressure.

Immersion liquid cooling adopts a completely different technical route: heating components such as chips are directly immersed in insulating cooling liquid, which has higher heat exchange efficiency, is easier to achieve extremely low PUE, and adapts to ultra-high power density computing clusters. The corresponding cost pressure comes from the systematic transformation of data center architecture, server hardware and operation and maintenance system.

The procurement and regular replacement of cooling liquid will bring continuous costs, and on-site operation and maintenance personnel also need to be retrained. This means that immersion liquid cooling is more suitable for large self-built clusters that operate stably for a long time and pursue extreme computing power density, and it is difficult to popularize it to general IDC scenarios in the short term.

Many manufacturers have made heavy bets on a single route to expand production based on their own judgment. Once downstream leading customers adjust their technical selection, the completed production lines will face the risk of capacity mismatch.

Compared with the route dispute between cold plate and immersion, the game between standardization and customization is more hidden. The data center solution in the air cooling stage has a high degree of standardization, and mass procurement can quickly replicate and implement.

However, most liquid cooling projects are customized projects, because the server models, cabinet layouts, and data center environments are different, and the parameters of pipelines, seals, and heat exchange need to be separately adapted and debugged. Even for the intelligent computing centers of the same customer in different parks, the solutions are different. The customized feature determines that it is difficult to expand production capacity as quickly as standardized hardware, which is the core reason why many manufacturers hold a large number of orders but encounter obstacles in delivery implementation.

Many new players in the market mistakenly believe that as long as they produce samples, they can take orders, underestimating the time and capital cost of repeated verification in multiple scenarios. Industry insiders believe that the two routes will still adapt to their corresponding scenarios in the short term, and it is difficult to tell the winner quickly. What determines the living space of enterprises is not only to choose the right technical route, but also the ability to continuously implement large-scale projects.

Computing power customers' parallel testing, long-term coexistence of technical routes and highly customized projects together form the complex background of the liquid cooling industry. After capital continues to pour in and a large number of new players enter the market, the focus of industry competition is shifting from simple technical solution comparison to the competition of supply chain access qualification and large-scale project delivery capability.

It should be noted that the core threshold of the track does not lie in basic manufacturing links such as sheet metal processing and pipeline assembly, but in the long and strict supply chain access of leading computing power customers, as well as the engineering capability of large-scale cluster delivery and long-term operation and maintenance.

The core contradiction of the current industry is very clear: letters of intent can be obtained through business negotiation, and factory production capacity can also be expanded rapidly with the help of capital, but the engineering capability to support large-scale stable delivery and the certification qualification of leading customers are difficult to establish in the short term.

Many manufacturers have sufficient orders on their books, but fall into the dilemma of "having orders difficult to deliver and idle production lines".

Final thinking on the dividend: Who can survive the cycle?

The current prosperity of the liquid cooling track is backed by the strong cyclical dividend brought by AI computing power construction.

Capital expenditure on computing power will not maintain a one-way continuous growth. The procurement plans of cloud vendors and intelligent computing operators will continue to dynamically adjust according to the implementation progress of large models, computing power utilization rate and budget status.

Once the deployment pace of computing power clusters slows down, the industry's demand expectation will adjust back quickly. In this environment, the first to bear the pressure are usually new players that choose asset-heavy expansion and only have letters of intent but have not yet entered the stable supply chain of leading customers.

Looking back at the cycles of hardware tracks such as energy storage and photovoltaics, we can find a highly similar development trajectory: in the demand explosion stage, capital enters the market quickly, and a large number of production capacities are implemented centrally; once the demand growth slows down and supply exceeds demand, a price war breaks out, and enterprises lacking core barriers are more likely to be squeezed out of the market.

Whether the liquid cooling track will repeat a similar cycle script depends on two core variables: first, the height of the supply chain certification barrier of leading customers, and second, the long-term maintained project customization level of the industry and the large-scale project delivery capability of manufacturers.

The customer access cycle of the liquid cooling track is longer and the engineering attribute is more prominent. The industry clearance may not be as rapid as that of the energy storage track, but the underlying logic of industry reshuffling is the same. Paper orders cannot resist cyclical fluctuations, and the real safety cushion for manufacturers comes from stable customer resources, mature processes and continuously iterative engineering capabilities.

First of all, deeply bind leading computing power customers and cut into their core supply chain system. The lengthy supplier certification itself is an industry barrier: once the large-scale project verification is completed, the replacement cost of customers will rise significantly, and manufacturers will have more opportunities to obtain deterministic orders, rather than only staying at the framework letter of intent stage.

Secondly, it is necessary to master core processes such as heat exchange, sealing, and material matching, and maintain the ability to continuously iterate. Manufacturers that only do OEM assembly can only earn processing margins during the industry dividend period; once downstream customers switch technical routes or put forward higher-density heat dissipation requirements, their own competitiveness will be easily impacted.

Furthermore, it is necessary to have large-scale cluster delivery and full-lifecycle operation and maintenance capabilities. The value of liquid cooling projects is not limited to hardware delivery. Leakage inspection, cooling liquid consumable replacement, and rapid fault response are all inseparable parts of the whole solution. Many problems in project implementation are not caused by defects in the equipment leaving the factory, but by the insufficient back-end operation and maintenance capability of manufacturers.

This also indicates that the concentration of the track is expected to continue to rise in the future period. Although the short-term order dividend attracts a large number of cross-border players to enter the market, as various pilot projects are implemented one after another, the shortcomings of manufacturers in reliability and engineering delivery will gradually be exposed.

The route differentiation between cold plate and immersion is still ongoing, the supporting industries such as upstream cooling liquid and sealing materials are still iterating, and customers' accumulation of long-term operation data of liquid cooling systems still requires a longer time. It can only be said that the current competition is still in the stage of engineering implementation verification, far from the time when the market pattern is finalized.

The "cool business" boom driven by the computing power wave is backed by the transformation of the underlying architecture of computing power infrastructure. The core test of the liquid cooling industry at present is not whether it can manufacture heat dissipation hardware, but to achieve long-term stable large-scale delivery in high-density computing power scenarios.

This liquid cooling boom driven by AI computing power is easily simplified as a hardware trend story: full orders, tight production capacity, and skyrocketing revenue. But peeling off the bustling order figures, the essence is the systematic reconstruction brought by the upgrading of computing power architecture. Heat dissipation is no longer an optional supporting facility in the data center, but the underlying base that determines whether high-density clusters can operate stably.

Computing power infrastructure itself comes with cycles, and AI capital expenditure will not always maintain the current rapid expansion. When the market heat cools down and the order dividend recedes, the screening mechanism of the track will really start.

The prosperity of the "cool business" is only a superficial phenomenon. The competition in the liquid cooling industry has just moved from sample testing and pilot projects to the core test of large-scale delivery.

This article is from the WeChat official account "Silicon Carbon Variable", author: Shen Lang, published with authorization from 36Kr.